Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 103
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
Claim(s) 1-4, 8-11, and 15-18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Haigh et al. (US Pub. No. 2018/0307945, published Oct. 2018, hereinafter "Haigh") in view of Olson et al. (NPL: TPOT: A Tree-based Pipeline Optimization Tool for Automating Machine Learning, published 2016, hereinafter "Olson"), and further in view of Feurer et al. (NPL: Auto-sklearn: Efficient and Robust Automated Machine Learning, published May 2019, hereinafter "Feurer").
Regarding claim 1, Haigh teaches a continuous machine learning system comprising:
a data generator module configured to obtain raw training data defining a total data size and to generate a plurality of data batches from the raw training data (Haigh, [0153] – “the AI system can include a training data loader 621 configured to load training data from a training data database 614a, a simulator 614b, and a streaming data server. The training data can be batched training data, streamed training data, or a combination thereof, and the AI engine can be configured to push or pull the training data from one or more training data sources selected from the simulator 614b, a training data generator, the training data database 614a, or a combination thereof. In an embodiment, a data stream manager can be configured to manage streaming of the streamed training data.” and in [0163] – “A batch data source can supply batched data from a database in at least one example.” – teaches a data generator module configured to obtain raw training data defining a total data size (size being the total amount of data received) and generating a plurality of data batches from the raw training data (data loader loads training data from a database or streaming data server. Training data can be batched training data. Batched training data is supplied from one data source or database));
to generate the refined pipeline in a containerized environment based on [a] best machine learning model pipeline, the refined pipeline configured to consume the plurality of data batches (Haigh, [0220] – “The set of independent processes, each independent process wrapped in its own software container, at least includes an instructor process and a learner process. The instructor process is configured to carry out a training plan codified in a pedagogical software programming language, and the learner process is configured to carry out an actual execution of underlying AI learning algorithms during a training session. The instructor process and the learner process of the set of independent processes cooperate with one or more data sources to train a new AI model.” – teaches to generate a refined pipeline (trained new AI model) in a containerized environment (each process wrapped in software container) based on the best machine learning model pipeline, the refined pipeline configured to consume the plurality of data batches (cooperates with one or more data source which can send data batches for refined pipeline to consume, as in [0163], therefore this can be considered a “best” pipeline as it is chosen to send));
a pipeline training module configured to incrementally train the refined pipeline using remaining data batches among the plurality of data batches generated after the initial data batch (Haigh, [0163] – “A batch data source can supply batched data from a database in at least one example.”, [0171] – “When in training mode the architect module 326 of the AI engine is configured to i) instantiate the network of processing nodes in any layers of hierarchy conforming to concepts of the problem being solved proposed by the user and ii) then the learner module 328 and instructor module 324 train the network of processing nodes in that AI model.” and in [0172] – “Because the process of building pedagogical programs is iterative, the AI engine in training mode can also provide incremental training.” – teaches a pipeline training module configured to incrementally trained the refined pipeline using remaining data batches among the plurality of data batches generated after the initial data batch (data source supplies batched data to system, architect module in training mode provides incremental training, thus incrementally training a refined pipeline using remaining data batches generated after the initial data batch)).
Haigh fails to explicitly teach a pipeline search module in signal communication with the data generator module to obtain an initial data batch from among the plurality of data batches and determine a best machine learning model pipeline among a plurality of machine learning model pipelines based on the initial data batch; a pipeline refinement module in signal communication with the pipeline search module to receive the best machine learning model pipeline, to refine the best machine learning model pipeline, and the best machine learning model pipeline that is selected based on an evaluation of multiple candidate machine learning model pipelines trained on the initial data batch.
However, analogous to the field of automated machine learning, Olson teaches:
a pipeline search module in signal communication with the data generator module to obtain an initial data batch from among the plurality of data batches and determine a best machine learning model pipeline among a plurality of machine learning model pipelines based on the initial data batch (Olson, Section 2.3 — “To begin, the GP algorithm generates 100 random tree-based pipelines and evaluates their balanced cross-validation accuracy on the data set.” and “at which point the algorithm selects the highest-accuracy pipeline from the Pareto front as the representative “best” pipeline from the run.” — teaches a pipeline search module in communication with the data generator to obtain data and determine a best machine learning model pipeline among a plurality of machine learning model pipeline based on the data (generates several pipelines and evaluates accuracy on provided data, selects highest accuracy pipeline as representative “best” pipeline));
the best machine learning model pipeline that is selected based on an evaluation of multiple candidate machine learning model pipelines trained on the initial data batch (Olson, Section 2.3 — “To begin, the GP algorithm generates 100 random tree-based pipelines and evaluates their balanced cross-validation accuracy on the data set.” and “at which point the algorithm selects the highest-accuracy pipeline from the Pareto front as the representative “best” pipeline from the run.” – teaches a best machine learning model pipeline that is selected based on an evaluation of multiple candidate pipelines trained on the initial data batch (100 random tree-based pipelines generated and evaluated, algorithm selects highest accuracy pipeline as representative best pipeline)),
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the plurality of data batches, incremental training, and containerized environment of Haigh to the model selection of Olson in order to search and select a best possible candidate machine learning model pipeline trained on an initial data batch. Doing so would enable the system to automatically design and optimize machine learning pipelines for a given problem domain (Olson, Section 1).
The combination of Haigh and Olson fails to explicitly teach a pipeline refinement module in signal communication with the pipeline search module to receive the best machine learning model pipeline and to refine the best machine learning model pipeline.
However, analogous to the field of automated machine learning, Feurer teaches:
a pipeline refinement module in signal communication with the pipeline search module to receive the best machine learning model pipeline, to refine the best machine learning model pipeline by replacing an initial estimator of the best machine learning model pipeline with a refined estimator (Feurer, Fig. 6.2 and Section 6.2 Paragraph 3 - “In a nutshell, Bayesian optimization [7] fits a probabilistic model to capture the relationship between hyperparameter settings and their measured performance; it then uses this model to select the most promising hyperparameter setting (trading off exploration of new parts of the space vs. exploitation in known good regions), evaluates that hyperparameter setting, updates the model with the result, and iterates.” – teaches a pipeline refinement module in communication with the search module to receive the best machine learning model pipeline and to refine the best machine learning model pipeline by replacing an initial estimator of the best machine learning model pipeline with a refined estimator (fits a probabilistic model to capture relationship between hyperparameter settings and measured performance, thus refining the received pipeline, by replacing the initial estimator of the model with refined estimator by selecting the most promising hyperparameter setting and updating model with the result));
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the pipeline refinement of Feurer to the data batches, incremental training, containerized environment, and model selection of Haigh and Olson to create a system that searches, refines, and trains a best machine learning model pipeline. Doing so would solve the fundamental problems of deciding which machine learning algorithm to use on a given dataset, whether and how to preprocess its features, and how to set all hyperparameters (Feurer, Introduction).
Regarding claim 8, Haigh teaches a computer-implemented method comprising:
obtaining, by a data generator module, raw training data defining a total data size; generating, by the data generator module, a plurality of data batches from the raw training data; (Haigh, [0153] – “the AI system can include a training data loader 621 configured to load training data from a training data database 614a, a simulator 614b, and a streaming data server. The training data can be batched training data, streamed training data, or a combination thereof, and the AI engine can be configured to push or pull the training data from one or more training data sources selected from the simulator 614b, a training data generator, the training data database 614a, or a combination thereof. In an embodiment, a data stream manager can be configured to manage streaming of the streamed training data.” and in [0163] – “A batch data source can supply batched data from a database in at least one example.” – teaches a data generator module configured to obtain raw training data defining a total data size (size being the total amount of data received) and generating a plurality of data batches from the raw training data (data loader loads training data from a database or streaming data server. Training data can be batched training data. Batched training data is supplied from one data source or database));
generating, based on the refining, the refined pipeline in a containerized environment based on the best machine learning model pipeline, the refined pipeline configured to consume the plurality of data batches (Haigh, [0220] – “The set of independent processes, each independent process wrapped in its own software container, at least includes an instructor process and a learner process. The instructor process is configured to carry out a training plan codified in a pedagogical software programming language, and the learner process is configured to carry out an actual execution of underlying AI learning algorithms during a training session. The instructor process and the learner process of the set of independent processes cooperate with one or more data sources to train a new AI model.” – teaches to generate a refined pipeline (trained new AI model) in a containerized environment (wrapped in software container) based on the best machine learning model pipeline, the refined pipeline configured to consume the plurality of data batches (cooperates with one or more data source which can send data batches for refined pipeline to consume, as in [0163]));
and incrementally training, by the pipeline training module, the refined pipeline using remaining data batches among the plurality of data batches generated after the initial data batch (Haigh, [0163] – “A batch data source can supply batched data from a database in at least one example.”, [0171] – “When in training mode the architect module 326 of the AI engine is configured to i) instantiate the network of processing nodes in any layers of hierarchy conforming to concepts of the problem being solved proposed by the user and ii) then the learner module 328 and instructor module 324 train the network of processing nodes in that AI model.” and in [0172] – “Because the process of building pedagogical programs is iterative, the AI engine in training mode can also provide incremental training.” – teaches a pipeline training module configured to incrementally trained the refined pipeline using remaining data batches among the plurality of data batches generated after the initial data batch (data source supplies batched data to system, architect module in training mode provides incremental training, thus incrementally training a refined pipeline using remaining data batches generated after the initial data batch)).
Haigh fails to explicitly teach obtaining, by a pipeline search module in signal communication with the data generator module, an initial data batch from among the plurality of data batches; determining, by the pipeline search module, a best machine learning model pipeline among a plurality of machine learning model pipelines based on the initial data batch; receiving, by a pipeline refinement module in signal communication with the pipeline search module, the best machine learning model pipeline; refining, by the pipeline refinement module, the best machine learning model pipeline; and the best machine learning model pipeline that is selected based on an evaluation of multiple candidate machine learning model pipelines trained on the initial data batch.
However, analogous to the field of automated machine learning, Olson teaches:
obtaining, by a pipeline search module in signal communication with the data generator module, an initial data batch from among the plurality of data batches (Olson, Section 2.2 Paragraph 1 — “where two copies of the data set are provided to the pipeline” – teaches the pipeline search module obtaining data);
determining, by the pipeline search module, a best machine learning model pipeline among a plurality of machine learning model pipelines based on the initial data batch (Olson, Section 2.3 — “To begin, the GP algorithm generates 100 random tree-based pipelines and evaluates their balanced cross-validation accuracy on the data set.” and “at which point the algorithm selects the highest-accuracy pipeline from the Pareto front as the representative “best” pipeline from the run.” — teaches determining an accuracy for each machine learning model pipeline based on the initial data, and determines the pipeline with the highest accuracy as the “best” machine learning model pipeline);
the best machine learning model pipeline that is selected based on an evaluation of multiple candidate machine learning model pipelines trained on the initial data batch (Olson, Section 2.3 — “To begin, the GP algorithm generates 100 random tree-based pipelines and evaluates their balanced cross-validation accuracy on the data set.” and “at which point the algorithm selects the highest-accuracy pipeline from the Pareto front as the representative “best” pipeline from the run.” – teaches a best machine learning model pipeline that is selected based on an evaluation of multiple candidate pipelines trained on the initial data batch (100 random tree-based pipelines generated and evaluated, algorithm selects highest accuracy pipeline as representative best pipeline)),
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the plurality of data batches, incremental training, and containerized environment of Haigh to the model selection of Olson in order to search and select a best possible candidate machine learning model pipeline trained on an initial data batch. Doing so would enable the system to automatically design and optimize machine learning pipelines for a given problem domain (Olson, Section 1).
The combination of Haigh and Olson fails to explicitly teach receiving, by a pipeline refinement module in signal communication with the pipeline search module, the best machine learning model pipeline; refining, by the pipeline refinement module, the best machine learning model pipeline;
However, analogous to the field of automated machine learning, Feurer teaches:
receiving, by a pipeline refinement module in signal communication with the pipeline search module, the best machine learning model pipeline (Feurer, Section 6.2 Paragraph 3 - “In a nutshell, Bayesian optimization [7] fits a probabilistic model to capture the relationship between hyperparameter settings and their measured performance;” – teaches receiving a machine learning model pipeline to be refined (fits a probabilistic model, thus receiving a best machine learning pipeline to be refined));
refining, by the pipeline refinement module, the best machine learning model pipeline by replacing an initial estimator of the best machine learning model pipeline with a refined estimator (Feurer, Fig. 6.2 and Section 6.2 Paragraph 3 - “In a nutshell, Bayesian optimization [7] fits a probabilistic model to capture the relationship between hyperparameter settings and their measured performance; it then uses this model to select the most promising hyperparameter setting (trading off exploration of new parts of the space vs. exploitation in known good regions), evaluates that hyperparameter setting, updates the model with the result, and iterates.” – teaches a pipeline refinement module in communication with the search module to receive the best machine learning model pipeline and to refine the best machine learning model pipeline by replacing an initial estimator of the best machine learning model pipeline with a refined estimator (fits a probabilistic model to capture relationship between hyperparameter settings and measured performance, thus refining the received pipeline, by replacing the initial estimator of the model with refined estimator by selecting the most promising hyperparameter setting and updating model with the result));
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the pipeline refinement of Feurer to the data batches, incremental training, containerized environment, and model selection of Haigh and Olson to create a system that searches, refines, and trains a best machine learning model pipeline. Doing so would solve the fundamental problems of deciding which machine learning algorithm to use on a given dataset, whether and how to preprocess its features, and how to set all hyperparameters (Feurer, Introduction).
Claim 15 incorporates substantively all the limitations of claim 8 in a computer readable storage medium, and is rejected on similar grounds as above.
Regarding claim 2, the combination of Haigh, Olson, and Feurer teaches the continuous machine learning system of claim 1, wherein refining the best machine learning model pipeline includes replacing the initial estimator of the best machine learning model pipeline with the refined estimator that includes the initial estimator along with one or more of an additional model hyperparameter (Feurer, Section 6.2 Paragraph 3 — “Bayesian optimization [7] fits a probabilistic model to capture the relationship between hyperparameter settings and their measured performance; it then uses this model to select the most promising hyperparameter setting (trading off exploration of new parts of the space vs. exploitation in known good regions), evaluates that hyperparameter setting, updates the model with the result, and iterates” — here, Feurer describes updating the initial estimator with the new hyperparameters resulting from Bayesian optimization, thus creating a refined estimator that includes the initial estimator and additional model hyperparameters that set a hyperparameter configuration on the refined pipeline.) and an additional feature engineering option (Feurer, Fig 6.2 — Figure 6.2 showcases feature pre-processing that includes Principle Component Analysis and fast Independent Component Analysis and Fig 6.2 showcases a data preprocessor that utilizes rescaling, one-hot encoder, data imputation, and balancing as feature engineering options), wherein the additional model hyperparameter sets a hyperparameter configuration on the refined pipeline, and the additional feature engineering option is configured to convert raw observations included in the initial data batch into one or more measurable inputs that correspond to desired features for use by the refined pipeline (Feurer, Section 6.2 Paragraph 3 — “Bayesian optimization [7] fits a probabilistic model to capture the relationship between hyperparameter settings and their measured performance; it then uses this model to select the most promising hyperparameter setting (trading off exploration of new parts of the space vs. exploitation in known good regions), evaluates that hyperparameter setting, updates the model with the result, and iterates” — teaches creating a refined estimator that includes the initial estimator and additional model hyperparameter that sets a hyperparameter configuration on the refined pipeline. Feurer further teaches in Fig. 6.2 and Section 6.4 Paragraph 3 – “They comprise data preprocessors (which change the feature values and are always used when they apply) and feature preprocessors (which change the actual set of features, and only one of which [or none] is used). Data preprocessing includes rescaling of the inputs, imputation of missing values, one-hot encoding and balancing of the target classes. The 14 possible feature preprocessing methods can be categorized into feature selection (2), kernel approximation (2), matrix decomposition (3), embeddings (1), feature clustering (1), polynomial feature expansion (1) and methods that use a classifier for feature selection (2). For example, L1-regularized linear SVMs fitted to the data can be used for feature selection by eliminating features corresponding to zero-valued model coefficients.” – teaches wherein the additional feature engineering option (in Fig. 6.2, see data preprocessing and feature preprocessing options) is configured to convert raw observations included in the initial data batch into one or more measurable inputs that correspond to desired features for use by the refined pipeline (as shown in Fig. 6.2 and explained in Section 6.4, feature preprocessing includes PCA, feature selection, clustering, embeddings, polynomial feature expansion, Fast ICA, e.g., PCA converts raw observations included in the data into one or more measurable inputs that correspond to desired features for use by the refined pipeline)).
Claims 9 and 16 are similar to claim 2, hence similarly rejected.
Regarding claim 3, the combination of Haigh, Olson, and Feurer teaches the continuous machine learning system of claim 2, wherein the pipeline search module is configured to determine an accuracy of each of the machine learning model pipelines (Olson, Section 2.3 — “To begin, the GP algorithm generates 100 random tree-based pipelines and evaluates their balanced cross-validation accuracy on the data set.” — Olson describes evaluating each machine learning model pipeline’s accuracy), and is configured to determine the machine learning model pipeline having a highest accuracy as the best machine learning model pipeline among the plurality of machine learning model pipelines (Olson, Section 2.3 — “at which point the algorithm selects the highest-accuracy pipeline from the Pareto front as the representative “best” pipeline from the run.” — Olson describes selecting the machine learning model pipeline having the highest accuracy as the best machine learning model pipeline among the plurality).
Claims 10 and 17 are similar to claim 3, hence similarly rejected.
Regarding claim 4, the combination of Haigh, Olson, and Feurer teaches the continuous machine learning system of claim 3, wherein the pipeline search module comprises:
an automated machine learning engine in signal communication with the data generator module and the automated estimator engine, the automated machine learning engine configured to create the plurality of machine learning model pipelines based on the initial data batch (Olson, Section 2.3 — “the GP algorithm generates 100 random tree-based pipelines and evaluates their balanced cross-validation accuracy on the data set” — teaches generating a plurality of machine learning model pipelines based on the initial data).
The combination of Haigh and Olson fails to explicitly teach an automated estimator engine configured to provide a plurality of different estimators, each estimator defined by one or a combination of a machine learning algorithm, an enhancement, and a feature engineering option;
However, analogous to the field of the claimed invention, Feurer teaches:
an automated estimator engine configured to provide a plurality of different estimators (Feurer, Section 6.3 Paragraph 1 - “Second, we include an automated ensemble construction step, allowing us to use all classifiers that were found by Bayesian optimization” — teaches an automated estimator engine that creates a plurality of estimators found by Bayesian optimization), each estimator defined by one or a combination of a machine learning algorithm, an enhancement, and a feature engineering option (Feurer, Section 6.2 — “Bayesian optimization [7] fits a probabilistic model to capture the relationship between hyperparameter settings and their measured performance; it then uses this model to select the most promising hyperparameter setting (trading off exploration of new parts of the space vs. exploitation in known good regions), evaluates that hyperparameter setting, updates the model with the result, and iterates” — describes the hyperparameter enhancement for a machine learning algorithm and in Fig 6.2 - showcases a data preprocessor that utilizes rescaling, one-hot encoder, data imputation, and balancing as feature engineering options); and
the plurality of different estimators (Feurer, Section 6.3 Paragraph 1 - “Second, we include an automated ensemble construction step, allowing us to use all classifiers that were found by Bayesian optimization” — teaches an automated estimator engine that creates a plurality of estimators (allows use of all classifiers found by Bayesian optimization))
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the plurality of estimators of Feurer to the data batches, pipeline search, and pipeline selection of Haigh and Olson in order to create a plurality of pipelines based on the initial data batch and the plurality of estimators. Doing so would solve the fundamental problems of deciding which machine learning algorithm to use on a given dataset, whether and how to preprocess its features, and how to set all hyperparameters (Feurer, Introduction).
Claims 11 and 18 are similar to claim 4, hence similarly rejected.
Claim(s) 5, 12, and 19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Haigh, Olson, and Feurer as applied to claims 1, 8, and 15 above, and further in view of Sakr et al. (EP Patent No. 3779806 Al (from IDS), published Feb. 2021, hereinafter "Sakr").
Regarding claim 5, the combination of Haigh, Olson, and Feurer teaches the continuous machine learning system of claim 4.
The combination of Haigh, Olson, and Feurer fails to teach wherein the plurality of machine learning model pipelines are assigned ranking scores in response to being trained according to the initial data batch, and wherein the pipeline search module selects the machine learning model pipelines having the highest ranking score as the best machine learning model pipeline and delivers the best machine learning model pipeline to the pipeline refinement module.
However, analogous to the field of automated machine learning, Sakr teaches:
wherein the plurality of machine learning model pipelines are assigned ranking scores in response to being trained according to the initial data batch (Sakr, [0019] — “The search optimizer 7 processes the data received from the knowledge base 9 to rank the identified machine learning pipelines based on their expected performance against the user-defined performance metric” — here, Sakr teaches assigning a ranking score to the machine learning model pipelines in response to being trained to an initial data batch), and wherein the pipeline search module selects the machine learning model pipelines having the highest ranking score as the best machine learning model pipeline and delivers the best machine learning model pipeline to the pipeline refinement module (Sakr, [0019] — “and selects candidate machine learning pipelines based on that ranking” — here, Sakr teaches selecting machine learning pipelines based on their ranking scores, with the machine learning pipeline with the highest ranking score being considered the “best”, and also in [0019] - “The search optimizer then sends data identifying the machine learning pipelines to an execution planner 11” — Sakr teaches delivering the best machine learning pipeline to another module).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the machine learning model pipeline ranking of Sakr to the system of Haigh, Olson, and Feurer to rank the machine learning model pipelines. Doing so would further improve the process of testing the performance for various different options for each stage of the machine learning pipeline to identify the best performing pipeline (Sakr, [0003]).
Claims 12 and 19 are similar to claim 5, hence similarly rejected.
Claim(s) 6 and 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Haigh, Olson, Feurer, and Sakr as applied to claims 1, 8, and 15 above, and further in view of Siebel et al. (WO2016/118979, published July 2016, hereinafter “Siebel”).
Regarding claim 6, the combination of Haigh, Olson, Feurer, and Sakr teaches the continuous machine learning system of claim 5.
The combination of Zhang, Olson, Feurer, and Sakr fails to explicitly teach wherein the data generator module loads a batch queue with a maximum number of data batches and continues to load the batch queue with data batches to maintain the maximum number of data batches, wherein the data generate loads the batch queue with a new data batch in response to outputting a loaded data batch from the batch queue to the pipeline search module.
However, analogous to the field of automated machine learning, Siebel teaches:
wherein the data generator module loads a batch queue with a maximum number of data batches and continues to load the batch queue with data batches to maintain the maximum number of data batches (Siebel, [0221] — “the continuous data processing component 1004 is configured to batch process data stress by the data services component 204, such as data in the one or more data stores 406-416 of FIG. 4.” — Siebel describes a continuous data processing in which data is batched continuously, and in [0233] — “For example, consumption of a stream processing analytic may be represented by messages of the data streams 1202 being fed and stored in one or more queues 1204” — Siebel teaches a stream of continuous data being stored in one or more queues that are constantly filled), wherein the data generate loads the batch queue with a new data batch in response to outputting a loaded data batch from the batch queue to the pipeline search module (Siebel, [0220] — “the continuous data processing component 1004 is configured to process data using one or more of Map reduce services, stream services, continuous analytics processing, and iterative processing” — Siebel teaches continuous data processing that involves loading the batch queue with a new batch in response to outputting a loaded data batch from the queue to another module, the data processing system of Siebel is continuous and configured to handle a large, constant stream of data from multiple data sources).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the continuous data batch queue of Siebel with the continuous machine learning system of Haigh, Olson, Feurer, and Sakr to create a system configured to take continuous batches of data to search and refine a machine learning model pipeline best fit to operate on the batches of data. Doing so would enable great flexibility based on the needs of a particular platform and may even improve machine learning speed and accuracy (Siebel, [0219]).
Claim 13 is similar to claim 6, hence similarly rejected.
Claim(s) 7, 14, and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Haigh, Olson, Feurer, Sakr, and Siebel as applied to claims 1, 8, and 15 above, and further in view of Zhang et al. (US Pub. No. 2021/0374602, published Dec. 2021, hereinafter “Zhang”).
Regarding claim 7, the combination of Haigh, Olson, Feurer, Sakr, and Siebel teaches the continuous machine learning system of claim 6,
The combination of Haigh, Olson, Feurer, Sakr, and Siebel fails to explicitly teach wherein incrementally training the refined pipeline includes generating a new version of the refined pipeline in response to training a current version of the refined pipeline using a most-recent data batch obtained from the batch queue.
However, analogous to the field of the claimed invention, Zhang teaches:
wherein incrementally training the refined pipeline includes generating a new version of the refined pipeline in response to training a current version of the refined pipeline using a most-recent data batch obtained from the batch queue (Zhang, [0030] — “After the models have been built, in some embodiments, they are evaluated or validated (e.g., using test data). In some embodiments, this can result in new Input Data 105 being acquired and prepared in order to refine the systems. Once the model(s) are acceptable, the process proceeds to Model Deployment 125, where they model(s) are used during runtime. In an embodiment, each element in the workflow may of course be repeated at any time. For example, after Model Deployment 125, the model(s) may be continuously or periodically refined using new data. Similarly, the models may be rebuilt completely (e.g., entirely retrained) at times.” — teaches incremental training (model continuously refined using new data) of the refined pipeline, where a new version of the pipeline is generated in response to training the current version using the most-recent data batch acquired (refined model is a new version of current model in response to being refined on most-recent data)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the generating a new version of the refined pipeline of Zhang to the batch data, incremental training, batch queue, and pipeline refinement of Haigh, Olson, Feurer, Sakr, and Siebel in order to generate a new version of the refined pipeline using a most-recent data batch. Doing so would increase the importance of optimized and efficient transformation pipelines, and enable continuous or periodic refinement of deployed pipelines in response to most recent data (Zhang, [0030]).
Claim 14 is similar to claim 7, hence similarly rejected.
Claim 20 is similar to claim 6 and 7, hence similarly rejected.
Response to Arguments
Applicant's arguments filed 20 March 2026 have been fully considered but they are not persuasive.
Applicant argues on pp. 2-3 that the cited portions of Haigh fails to teach the limitations reciting an initial data batch and remaining data batches from the plurality of data batches generated from raw training data defining a total data size. Applicant further argues that the data described by Haigh corresponds to data that is supplied or processed without being generated as a plurality of data batches partitioned into an initial data batch and remaining data batches. Examiner respectfully disagrees. As Haigh states in [0153] – “In addition to the foregoing, the AI system can include a training data loader 621 configured to load training data from a training data database 614a, a simulator 614b, and a streaming data server. The training data can be batched training data, streamed training data, or a combination thereof, and the AI engine can be configured to push or pull the training data from one or more training data sources selected from the simulator 614b, a training data generator, the training data database 614a, or a combination thereof. In an embodiment, a data stream manager can be configured to manage streaming of the streamed training data.” and in [0163] – “The instructor module 324 may flow data to train AI objects from many data sources including, but not limited to a simulator, a batch data source, a random-data generator, and historical/guided performance form from past performance… A batch data source can supply batched data from a database in at least one example.” – which states that a data loader is configured to load training data from a training data database, in which the training data is batched training data, streamed training data, or a combination thereof. Thus, the first data loaded by the data loader 621 is considered as the initial data batch, and any data batch remaining thereafter is the remaining data batches within the batch loader. Additionally, Haigh states in [0162] – “The learner module 328 and instructor module 324 can work with a simulator or other data source to iteratively train an AI object with different data inputs. The instructor module 324 can reference a knowledge base of how to train an AI object efficiently by different ways of flowing data to one or more AI objects in the topology graph in parallel, or, if dependencies exist, the instructor module 324 can train serially with some portions of lessons taking place only after earlier dependencies have been satisfied.” – which corresponds to different ways of flowing data to an AI, such that the instructor module can train serially with some portions of training taking place only after earlier dependencies are satisfied. Thus, the instructor module can utilize an initial data batch in, for example, Olson’s pipeline search and incrementally train (Haigh teaches incremental training at [0172]) the refined pipeline using remaining data batches thereafter according to data flow.
Applicant argues on pp. 4-5 of Remarks that Feurer fails to teach receiving a best machine learning model pipeline and replacing an initial estimator of the best machine learning pipeline with a refined estimator. Examiner respectfully disagrees. Feurer teaches at Section 6.2 Paragraph 3 - “In a nutshell, Bayesian optimization [7] fits a probabilistic model to capture the relationship between hyperparameter settings and their measured performance;” – fitting a probabilistic model to capture relationship between hyperparameter settings and measured performance, the probabilistic model includes the best machine learning model pipeline searched by Olson. Further, in Fig. 6.2 description – “Squared boxes denote parent hyperparameters whereas boxes with rounded edges are leaf hyperparameters. Grey colored boxes mark active hyperparameters which form an example configuration and machine learning pipeline. Each pipeline comprises one feature preprocessor, classifier and up to three data preprocessor methods plus respective hyperparameters” and Section 6.2 Paragraph 3 - “In a nutshell, Bayesian optimization [7] fits a probabilistic model to capture the relationship between hyperparameter settings and their measured performance; it then uses this model to select the most promising hyperparameter setting (trading off exploration of new parts of the space vs. exploitation in known good regions), evaluates that hyperparameter setting, updates the model with the result, and iterates.” – which states using the model to select the most promising hyperparameter setting, and updating the model with the result, which teaches replacing the initial estimator with a refined estimator by taking the model and performing hyperparameter optimization, thus replacing the initial estimator with an optimized estimator. In regards to applicant’s arguments on pp. 5 of Remarks, the cited passages of Feurer above teach modifying a machine learning pipeline such that the refined estimator includes the initial estimator with additional model hyperparameters and an additional feature engineering option (See Fig. 6.2 & See revised 35 U.S.C. 103 rejection above).
In response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Zeng et al. (NPL: Progressive sampling-based Bayesian optimization for efficient and automatic machine learning model selection, published Sept. 2017) teaches a progressive sampling-based Bayesian optimization for automatic selection for algorithms and hyperparameter values. Teaches using an initial sample from a raw training data set to perform a first round, which includes using the initial sample to test various combinations of machine learning models and hyperparameter configurations and quickly removing unpromising models. Teaches subsequent rounds that further train the best model using the remaining samples after the initial sample.
Crankshaw et al. (NPL: InferLine: ML Inference Pipeline Composition Framework, published Dec. 2018) teaches a system which efficiently provisions and executes ML inference pipelines subject to end-to-end latency constraints, with proactive optimization and control of model configuration. Teaches receiving a raw training data set that is partitioned into batches, buffered in the form of queues between stages of the pipeline.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LOUIS C NYE whose telephone number is 571-272-0636. The examiner can normally be reached Monday - Friday 9:00AM - 5:00PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, MATT ELL can be reached at 571-270-3264. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/LOUIS CHRISTOPHER NYE/Examiner, Art Unit 2141
/TAN H TRAN/Primary Examiner, Art Unit 2141
Prosecution Insights